An automated LP agent for Meteora DLMM. Built, tested, and learned why absolute return is hard.
| Aspect | Description |
|---|---|
| Goal | Explore automated LP strategies for absolute return |
| Status | Research / Learning (not live trading) |
| Pair | SOL/USDC (Meteora DLMM) |
| Period | Dec 2024 – May 2026 (SOL $261 → $78, -70%) |
| Metric | Best Strategy (30% exit / 168h delay) |
|---|---|
| LP Return | -11.47% |
| HOLD Return | -46.68% |
| Outperform | +35.21% |
| Exits | 6 |
| Gas Cost | $1.80 (realistic $0.15/tx) |
Conclusion: No strategy achieved absolute profit in this bear market, but exit strategy significantly minimized loss.
- Backend: Node.js, PM2, Solana Web3.js
- Infra: VPS (Ubuntu), Helius RPC + Fallback
- Monitoring: Telegram bot, Web Dashboard
- Backtesting: Python, pandas, real Binance data
This project was built using AI-as-Engineer methodology, not "vibe coding" or random prompting.
| Aspect | Approach |
|---|---|
| Strategic Direction | Problem framing, hypothesis generation, experiment design |
| AI Management | Directed multiple AI models (DeepSeek as lead architect, Claude, Gemini, Kimi for expert validation) |
| Quality Control | Reviewing outputs, rejecting bad code, maintaining consistency |
| Integration | Merging AI-generated modules into production-ready system |
| AI | Role | Cost |
|---|---|---|
| DeepSeek | Lead Tech Architect + Strategic Advisor | $0 |
| Claude | DeFi Expert (brutal analysis) | $0 |
| Gemini | Strategy Consultant | $0 |
| Kimi | DLMM Specialist | $0 |
- Not "prompt and pray" — Every AI output was reviewed, tested, and integrated intentionally
- Cost-effective — $0 infrastructure for ideation to execution
- Reproducible — Method can be applied to any DeFi strategy exploration
- Transparent — Full development log shows AI collaboration process
"I don't code. I orchestrate AIs who code."
This project is a case study in AI-augmented engineering — leveraging LLMs as domain experts, code generators, and reviewers while maintaining human strategic control.
📖 See DEVELOPMENT_LOG.md for the full AI collaboration timeline.
- DLMM on volatile pairs is better for DCA than absolute return
- Exit strategy (30%/168h) outperforms no-exit by +11%
- Gas optimization is secondary to strategy viability
- Knowing when to stop is a feature, not a bug
📖 Full Lessons Learned
gtrade/ ├── src/ # Engine modules (reusable) ├── scripts/ # Backtesting & data fetching ├── data/ # Historical candles ├── docs/ # Blueprint, lessons, AI insights ├── DEVELOPMENT_LOG.md └── README.md
- Research alternative pools: USDC/USDT, SOL/JitoSOL
- Apply same framework to correlated pairs
- Backtest for absolute positive return
Expert insights from Claude, Gemini, and Kimi (DeFi LP specialists).
Built with curiosity and rigor.
Not every experiment yields profit. But every experiment yields knowledge.